Scenario Planning
for Generative AI

The staircase pattern

Looking backward, AI capability has advanced in a staircase pattern: a major jump every roughly two years, followed by refinement within that generation. GPT-3 (2020) was dramatically surpassed by GPT-4 (2023), which was then refined through GPT-4 Turbo, GPT-4o, and eventually GPT-4o-mini, each iteration better and cheaper within the same capability tier. The same pattern appears with Claude 3 Opus giving way to Claude 3.5 Sonnet, then Claude Opus 4, refined through Opus 4.6 and 4.7. The next step, the Mythos generation, is due on this rhythm toward the end of 2026 and into 2027. What landed in June 2026 was an early variant of it; more on that below.

If this pattern holds, the $228 billion spent in 2024 is currently producing the training infrastructure for models that ship in 2026–2027. The $416 billion committed for 2025 funds models arriving in 2027–2028. And the ~$700 billion planned for 2026 is investing in capabilities whose architectures may not even be designed yet: research compute for ideas that haven’t been conceived. The clearest demonstration of this lag came on March 24, 2026, when OpenAI completed pre-training of GPT-6 (“Spud”) at the Abilene Stargate facility, a model whose existence was funded by 2024’s capex, roughly three years before its expected public launch. Anthropic’s next-generation “Mythos” sits in the same lane: limited testing with cybersecurity defenders through Q1 2026, and then, on June 9, 2026, a safeguard-wrapped variant shipped publicly as Claude Fable 5 (the unrestricted Mythos 5 stayed reserved for cyberdefenders), priced at $10 input / $50 output per million tokens. Capability paid for by money committed before its architecture had a name.

That June release is why Fable 5 carries an asterisk on the timeline above, plotted without a parameter count. Our readinginterpretation: the staircase thesis holds, with the full Mythos / Spud generation remaining on the 2027 mark exactly where the lag pipeline puts it, and Fable 5 looks like that generation released early, in a deliberately wrapped form. The shape of the release supports this. Instead of the final rounds of fine-tuning, the alignment and moderation work appears to ride on a software layer beside the model (conservative safeguards that route sensitive topics to Opus 4.8), and the timing, eight days after Anthropic’s S-1 filing, suggests the IPO calendar may have pulled the launch forward. Treat the early arrival as a preview of the step rather than the step itself: the trigger to watch is the full deployment.

Postscript, July 2026: three days after launch, a federal directive took both models dark worldwide; they returned gated, with Mythos flowing through Project Glasswing for large US organizations and Fable available via API tokens behind a strong moderation layer, at premium prices. The capability reading above is unchanged: the step previewed is real, and the Remote Labor Index measured it before the restrictions landed (Current 6). But the access story has become its own force. Current 4 picks it up, and the companion booklet gives it a full treatment.

How to use it from here

The decoder gives you a discipline for reading AI news. When a capability headline lands, ask which year’s money produced it. When a capex announcement lands, ask which years it will produce: serving capacity in two, models in three, breakthroughs in four. The eight currents that follow are the forces pulling on this machinery: whether the staircase keeps climbing (Current 1), whether efficiency strands the clusters (Current 2), whether the investment timeline survives contact with the revenue timeline (Current 3), whether a sovereign cuts the wire between you and your vendor (Current 4), whether any of it actually reaches production (Current 5), whether the hours-and-dollars math starts moving work (Current 6), whether the atoms of power, chips, and permits keep pace with the money (Current 7), and what politics does when the displacement math starts working (Current 8). Each current ends with trigger signals to watch, and the decoder is how you date a trigger when it fires.

The bet, stated plainly

The capex decoder shows where the money goes and when it comes back as capability. This current is the bet that the spending keeps being right: that above the step the field just took, there is at least one more, and that the clusters being poured today will be the ones that train it.

Note what just happened to this bet: it began paying out. The 2024 cohort of capex was earmarked, in part, for models shipping 2026–2027, and on June 9, 2026, an early variant of the Mythos generation arrived as Claude Fable 5, with the full generation (Mythos proper, OpenAI’s Spud) still expected toward late 2026 and into 2027, right on the lag pipeline’s schedule. The live question is therefore twofold: whether the full deployment confirms the step the early variant previews, and whether the ~$700 billion of 2026 money buys a step above it, or the staircase flattens just as the most expensive clusters in history come online. That is what the trigger signals below are tuned to detect.

What is being built

The scale of individual projects is staggering. Elon Musk’s xAI deployed Colossus, a cluster of 100,000 GPUs, in just 122 days in late 2024. Project Rainier, the cluster Amazon built for Anthropic, spans roughly 500,000 chips. The Stargate project (a joint venture between OpenAI, Oracle, and SoftBank) targets 450,000 GPUs at its Abilene, Texas flagship and plans clusters exceeding one gigawatt of power, the equivalent of a nuclear power plant dedicated to AI, while Meta’s Hyperion campus in Louisiana is designed for five. At these scales, the limiting factor shifts from chip availability to raw electrical power: the Three Mile Island nuclear facility is being restarted specifically to supply AI data centers. Whether the atoms can keep pace with the money is its own force now; Current 7 takes up that story in full.

The inference bet

A common misunderstanding is that all this money is about training bigger models. In reality, the labs are increasingly betting that inference, running models at scale for millions of users, will dominate AI compute. Deloitte’s 2026 predictions already put inference at roughly two-thirds of all AI compute this year reported, and Brookfield’s infrastructure forecast has it at roughly 75% of AI compute demand by 2030 reported. Training a frontier model is a one-time cost; serving it to every enterprise customer, every developer, every consumer product is a continuous cost that scales with adoption. The capex surge is as much about building the serving infrastructure for AI-powered products as it is about training the next generation of models.

This is the core planning insight of Current 1: even if you believe the current generation of AI is “good enough,” the investment already committed will produce outcomes over the next 2–4 years. Those outcomes (faster models, cheaper inference, new capabilities) will change what’s possible and what’s expected. Your plans need to account for a moving target rather than a snapshot.

The training cost freefall

In March 2023, OpenAI trained GPT-4 for an estimated $63–100 million; Sam Altman confirmed publicly that the cost exceeded $100 million including research and development. By July 2024, Meta had trained Llama 3.1 405B, an open-weight model matching GPT-4 on most benchmarks, for roughly $60 million in compute (30.84 million H100 GPU-hours). Then in December 2024, DeepSeek released V3, a model that matched or exceeded GPT-4o on key benchmarks for just $5.6 million in GPU time.

That is a roughly 95% cost reduction in 20 months. A month later, DeepSeek released R1, which matched OpenAI’s o1 on reasoning tasks for an incremental $294,000 in training cost.

The caveats matter: DeepSeek’s $5.6 million figure covers only the final pre-training run. Their parent company High-Flyer invested over $500 million in Nvidia GPUs total, and the full cost from base model to R1 is estimated at $6–7 million by Epoch AI. But even the generous estimate represents a 90%+ reduction from GPT-4. The key innovations enabling this (FP8 mixed-precision training, mixture-of-experts architectures with load balancing, and custom CUDA kernels achieving 85%+ GPU utilization versus the industry average of 55–65%) are algorithmic rather than hardware-bound. They can be replicated.

The open-source convergence

The Stanford HAI 2025 AI Index Report documented the most important shift in the AI landscape: the performance gap between the best open-weight and proprietary models, measured by Chatbot Arena Elo ratings, shrank from 8.04% in January 2024 to 1.7% by February 2025, a 79% reduction in a single year. On MMLU specifically, the gap between US and Chinese models collapsed from 17.5 percentage points to just 0.3 between the end of 2023 and the end of 2024.

Llama 3.1 405B was the first open model to match or exceed GPT-4 across multiple benchmarks in July 2024, roughly 16 months after GPT-4’s release. By early 2025, that lag had compressed further. Open-source models now represent 62.8% of all models by count, and the best open LLMs lag closed ones by 5–22 months on benchmarks. One analysis projected open-closed parity by Q2 2026; that projection has aged poorly at the very top, where the 2026 AI Index shows the Elo gap re-widening to roughly 3.3% after the frontier stepped again. Read the convergence as a per-tier story: each capability tier commoditizes on the 5–22-month lag, while a genuinely new tier can re-open the gap above it.

Inference pricing in freefall

The pricing evolution of OpenAI’s own API tells the commoditization story in dollar terms. GPT-4 launched at $60 per million output tokens in March 2023. GPT-4 Turbo brought that down to $30 in November 2023. GPT-4o launched at $15 in May 2024, then was cut to $10 in October. Meanwhile, GPT-4o-mini offered GPT-4-class performance at $0.60 per million tokens, a 99% reduction from GPT-4’s launch price in under two years.

Open-source alternatives are even cheaper. Llama 3 70B via Groq costs roughly $0.79 per million output tokens. DeepSeek V3 is available at $0.42. Self-hosted 70B models on H100 hardware can reach approximately $0.07 per million tokens at full utilization. On average, open-source models cost 7.3 times less than their proprietary equivalents.

The April 2026 wave

Over an 18-day window in April 2026, three frontier-class open-weight coding models shipped. GLM-5.1 (Zhipu, April 7) is 744B MoE / 40B active under MIT license and posts 58.4% on SWE-Bench Pro, ahead of GPT-5.4 and Opus 4.6 on that bench. Qwen 3.6 (Alibaba, April 16) split into variants; the 35B-A3B open variant runs on a single RTX 4090 with quantization and posts 73.4% SWE-Bench Verified, the throughput sweet spot for solo developers and on-prem deployments. Kimi K2.6 (Moonshot, April 20) is 1T total / 32B active and introduces a 300-agent swarm primitive for parallel exploration on hard tickets, an agentic precursor that connects to Hours and Dollars in Section 6. DeepSeek V4-Pro (April 24) posts 80.6% SWE-Bench Verified at $0.28 input / $2.48 output per million tokens with a 1M-token context window, roughly an order of magnitude cheaper than Opus 4.6 at comparable coding capability, using hybrid attention (Compressed Sparse + Heavily Compressed) at about 27% of V3.2’s per-token FLOPs. Mistral Medium 3.5 (April 29) ships as a 128B dense model, self-hostable from Hugging Face, with 77.6% SWE-Bench Verified and configurable reasoning effort per request; it carries the cleanest EU-data-residency narrative on the market, paired with Mistral’s $400M ARR (January 2026) at a $13.8B valuation.

By mid-2026, the buyer question is no longer “is open-weight good enough.” It is which open-weight per workload, which hosting stack, and which sovereign deployment shape. For EU enterprises in particular, these intersect directly with the sovereignty story in Section 4: for the first time, “self-hostable frontier-equivalent” is a literal product description rather than a euphemism.

Where does value go when the model is free?

Mistral CEO Arthur Mensch has been the most articulate voice on this shift. Across early-2026 interviews (the Big Technology Podcast in January, Davos and Bloomberg the same month, the Economic Times in February) he framed AI as becoming infrastructure, “a utility” measured by efficiency, capital discipline, and reliable delivery rather than novelty. His most quoted line: “My generation of engineers has more or less succeeded in commoditizing its own profession.” The corollary, which he argues consistently, is that competitive advantage will increasingly accrue to whoever builds the most specialized system around their specific data and domain, rather than to whoever has the largest model.

If Mistral is right, and the cost data supports the argument, then the model itself becomes a commodity layer, and value migrates to the layers around it: fine-tuning and domain adaptation, data pipelines and retrieval-augmented generation, tooling and orchestration (agent frameworks, MCP servers, evaluation pipelines), and ultimately domain expertise. The organizations that win in this scenario are the ones that understand their own problems best, whichever model they happen to run.

The dot-com precedent

On March 10, 2000, the Nasdaq Composite reached an all-time high of 5,048.62. By October 9, 2002, it had fallen to 1,114, a 78% decline that destroyed over $5 trillion in market value. The Nasdaq didn’t close above 5,000 again until April 23, 2015, a recovery that took fifteen years. At the peak, venture capital investment had surged from roughly $7 billion in 1995 to nearly $100 billion in 2000, with internet companies absorbing 80% of all venture capital. Telecom companies invested more than $500 billion in infrastructure in the five years following the 1996 Telecommunications Act.

The lesson that most people take from this period is: “it was a bubble and it burst.” The more useful lesson is that the survivors and casualties were distinguished by one thing: real revenue, real customers, and cash to survive a funding drought. The quality of their technology mattered surprisingly little.

The bubble argument has matured

The most visible bear voice of the past two years, Ed Zitron, has been notable not for being right but for how the argument has had to shift. His original case, sustained across blog posts and his “Better Offline” podcast, was economic: AI was a value-destruction machine, hyperscaler capex was insane, and the unit economics simply did not work. Some of that case has aged well (the ROI gap is real). Most of it has not. Frontier-tier inference prices fell roughly 83% in 19 months (99% if you accept a cheaper capability tier, a distinction Current 2 unpacks). Anthropic’s annualized revenue passed OpenAI’s in April 2026 ($30B vs $25B)reported and reached a reported $47B run-rate by mid-May. Cost decline plus revenue growth made the original economic argument harder to sustain in its strongest form. Kelsey Piper, writing in The Argument, documented the shift: Zitron’s case has migrated from “the economics don’t work” toward fraud and accounting allegations against OpenAI and the hyperscalers.

The bear case is still alive, and parts of it remain sharp. But the goalposts moved, and that itself is a signal worth weighing. A bubble argument that survives collapsing costs and compounding revenue by switching from economics to fraud is a weaker argument than one that didn’t have to switch. Hold the correction scenario open; don’t hold this particular version of it as the bear case.

Amazon vs. Pets.com

Amazon’s stock fell 94% from roughly $106 in December 1999 to about $5.51 in late 2001. Yet its revenue grew every single year through the crash: $2.76 billion in 2000, $3.12 billion in 2001, $5.26 billion in 2003, $8.49 billion in 2005. It posted its first profitable quarter in Q4 2001 and its first full profitable year in 2003, with $35 million net income on $5.26 billion revenue. The key decision was a well-timed $1.25 billion bond offering that gave Amazon $1 billion in cash to survive the drought. Today it is worth roughly $2.5 trillion, over 800 times its trough market cap.

Pets.com raised $300 million total, spent over $70 million on advertising while generating only $619,000 in revenue, and shut down 268 days after its IPO. Webvan burned through $1.5 billion building automated warehouses before filing bankruptcy. Boo.com raised $135 million, burned it in 18 months, and sold its assets for under $2 million. The common thread: negative unit economics, no path to profitability, and complete dependence on the next funding round.

AI investment has entered unprecedented territory

The combined Big Four capex (Alphabet, Amazon, Meta, Microsoft) grew from roughly $140 billion in 2023 to $228 billion in 2024 (+62% year-over-year), to $416 billion in 2025, to roughly $700 billion guided for 2026, the same series the capex decoder plots. Capital intensity has reached historically unprecedented shares of revenue for these companies. Venture funding has concentrated similarly: global AI VC funding grew from roughly $45–50 billion in 2022 to $211 billion in 2025, the first year AI startups captured more than half (52.7%) of all global venture deal value.

OpenAI reached an $852 billion post-money valuationreported after its $122 billion funding round in March 2026. Annualized revenue hit $25 billion by February 2026reported, up from roughly $2 billion in 2023. But the company projects a $14–17 billion loss in 2026projected, is not expected to be profitable until 2029 at the earliest, and has committed $600 billion in compute spending through 2030. Anthropic’s trajectory has been even steeper: revenue grew from $1 billion ARR in December 2024 to roughly $9 billion at the end of 2025 to a $47 billion run-rate by mid-May 2026reported, its Series H closed at $965 billion post-moneyreported, and on June 1, 2026 it filed a confidential S-1, the most concrete IPO signal short of a roadshow.

The revenue gap

Sequoia partner David Cahn published “AI’s $600B Question” in June 2024, calculating that the AI infrastructure buildout requires roughly $600 billion in annual end-user revenue to justify itself. At the time, actual AI product revenue was roughly $100 billion, leaving a $500 billion annual gap. Since then, both spending and revenue have grown, but spending has grown far faster: capex roughly tripled while the revenue gap has likely widened, not narrowed. Barclays estimated that current capex levels would require the equivalent of 12,000 ChatGPT-sized products to break even.

Personal value is clear. Enterprise value is the open question.

Two ROI stories sit on top of each other and are routinely conflated. Individual subscribers buying Claude or ChatGPT at $20–$200 a month report value clearly and stickily: paid consumer plans for the two leaders together cross tens of millions of seats by mid-2026, churn is unremarkable, and surveys consistently show personal users describing meaningful time savings. That part of the market has answered. The enterprise market has not.

Omdia’s October 2025 survey of 350 mid-to-large enterprises reported “very good” to “extraordinary” ROI from most respondents, a genuinely positive signal. Accenture, in parallel, found 61% of enterprise AI subscriptions underutilized due to poor integration. The MIT NANDA study reported 95% of organizations seeing zero return, with the measurement caveats already noted (no baselines, six-month cutoff, parallel-pilot designs). Reconcile these and the picture is: enterprises that have integrated AI into workflows are extracting real value; the majority that are still trying are not. The model can usually do the task, so the gap has little to do with capability. The bottleneck is how the model gets wired into the workflow. That is the subject of Section 5.

Vendor concentration is the under-discussed risk

Q1 2026 saw AI venture funding concentrate to a degree that has no recent precedent in software. OpenAI, Anthropic, and xAI accounted for roughly 67.3% of all AI venture funding across more than 1,500 deals. OpenAI’s $122 billion round at an $852 billion valuation consumed a non-trivial share of global venture capacity. Microsoft, Meta, Amazon, and Alphabet collectively guided investors toward ~$700 billion of capex in 2026. Three foundation-model labs sit on top of a stack the rest of the industry rents from.

Concentration this extreme is usually argued as safety: the giants won’t fail. The Anthropic-Pentagon situation (covered in Section 4) is the case to study before agreeing. A single sovereign decision in February 2026 severed access to a major AI vendor for the entire US federal government, mid-contract, with little notice. The technology kept working. The vendor didn’t fail. The buyer simply couldn’t buy. June 12 then supplied the general case: one directive, and every user of a model lost it at once (Current 4). That is a vendor-concentration failure mode the dot-com analogy didn’t have. Stress-testing your AI strategy against vendor severance is now first-class planning work, not paranoia.

Why the parallel breaks, and why it might not matter

There are important differences from the dot-com era. Today’s leading AI investors are massively profitable companies spending from earnings, not startups burning venture capital. Nasdaq forward price-to-earnings ratios are approximately 26 times versus 60 times at the dot-com peak. Enterprise adoption is far more advanced: the large majority of big enterprises have implemented AI in some form, even if mostly in pilots. But the core structural risk, investment dramatically outpacing revenue realization, is identical. And new risks have emerged: AI-related corporate debt has ballooned to $1.2 trillion (JPMorgan), GPU rental prices have already fallen roughly 70% from peak, and the real useful life of GPU infrastructure may be 2–3 years rather than the 5–6 years used for accounting depreciation.

AI is valuable; that part is settled. The planning question is whether your specific vendors, tools, and providers are the Amazon or the Pets.com of this cycle.

Two collisions, one pattern

Two events in spring 2026 turned sovereignty from hypothetical to operational. The first: on February 27, 2026, the US Defense Secretary designated Anthropic a “supply chain risk,” and the Trump administration ordered federal agencies to stop using Claude. Anthropic and the Pentagon had signed a $200 million contract in July 2025 under Anthropic’s acceptable-use policy; the Pentagon wanted “all lawful purposes” access without limitation, and Anthropic refused to remove restrictions on autonomous weapons and domestic mass surveillance. Anthropic sued, won a preliminary injunction in late March (Judge Rita Lin called the designation “Orwellian” and First Amendment retaliation), then lost an appeals court bid in early April. As of June 2026 the litigation is still live, and the designation has proven narrower in practice than first implied: Anthropic argued, and Microsoft agreed, that it cannot reach customers outside the defense contracts themselves, so Claude remained available through M365, GitHub, and AI Foundry. But the headline lesson stands: the technology never stopped working. The buyer simply could not buy.

The second: on April 8, Meta launched Muse Spark, its first proprietary closed-weight model, from Meta Superintelligence Labs under Alexandr Wang. After nearly a decade of public commitment to open frontier AI, Meta’s frontier development is now closed. Existing Llama models remain available but no longer evolve. The combination of $115–135B in 2026 capex, competitive pressure from Chinese labs building commercial products on top of Llama, and the strategic goal of a deeply integrated “personal superintelligence” tied to Meta’s user data drove the shift. Yann LeCun, Meta’s most visible open-source advocate, departed in November 2025. The “Linux of AI” thesis did not survive contact with $100B+ compute economics, at least at the Western frontier.

June 12: the collision generalized

Then, days before this edition closed, the pattern escalated from procurement to existence. On June 12, 2026, three days after the Fable 5 launch, the US government had Anthropic switch off Claude Mythos 5 and Fable 5 for all users worldwide, under a directive scoped by foreign nationality that, being unverifiable in practice, forced a global takedown. By early July the capability was coming back, but gated and priced: full Mythos access limited to large US organizations through Project Glasswing; Fable restored only through API tokens, at premium prices, behind a strong moderation layer. February’s lesson was that one buyer can lose its vendor. June’s lesson is stronger: every buyer can lose the model at once, and access comes back, when it comes back, on the sovereign’s terms, tier by tier. The companion booklet was written in the hours after that morning; it treats June 12 not as an outage but as a dress rehearsal.

The Chinese open-weight wave fills the gap

Within 18 days in April 2026, three frontier-class open-weight coding models shipped from Chinese labs: GLM-5.1 (Zhipu, MIT-licensed 744B MoE / 40B active, 58.4% SWE-Bench Pro), Kimi K2.6 (Moonshot, 1T total / 32B active with a 300-agent swarm primitive), and DeepSeek V4-Pro (1.6T total / 49B active, 80.6% SWE-Bench Verified at roughly an order of magnitude lower output cost than Opus 4.6). Qwen 3.6 (Alibaba, April 16) split into variants; the 35B-A3B open variant runs on a single RTX 4090 with quantization. By workload as of mid-2026: DeepSeek V4-Pro for cheap large-context coding agents, Kimi K2.6 for hard multi-step tickets with its swarm primitive, GLM-5.1 for self-hosted production where MIT licensing matters, Qwen 3.6-35B-A3B for local laptop or single-GPU deployment. Llama 4 remains integration-default but is no longer evolving.

What this means for EU enterprises

The buyer question has shifted from “is open-weight good enough” to a multi-part procurement question: which open-weight per workload, which hosting stack, which sovereign deployment shape. Mistral Medium 3.5 (128B dense, self-hostable, $400M ARR at a $13.8B valuation) is the cleanest EU-data-residency narrative on the market: dense rather than MoE, easier to deploy than the Chinese stacks, sovereign-aligned. Self-hosted Chinese open-weight is the other major option, with two caveats worth flagging: geopolitical exposure if procurement frameworks tighten, and dataset-provenance questions that some EU regulators are starting to ask.

The deeper point: for the first time in this booklet’s lifetime, “self-hostable frontier-equivalent” is not a euphemism. Procurement teams that previously assumed one or two US vendors had no realistic alternative now have several. The work shifts from negotiating with one vendor to architecting around the choice. That choice depends on which sovereign failure modes you weight highest.

Where this current continues

Sovereignty is the one current where this booklet deliberately caps its own depth: the companion booklet, The Mercantilism of Generative AI, treats it at book length. If you take two pieces of it, take these. First, the armed-versus-targeted test: which industries get frontier access on favorable terms because the factory-holder can never become them, and which get the bloc’s champion pointed at their customers because their margin is the cognitive work the factory produces. Second, the “open is a position, not a principle” mechanism: the reason this chapter prices the Chinese open-weight option as a strategy that can change rather than a guarantee, and the source of the cleanest single tell in either booklet: the day a leading Chinese lab ships its best model closed is the day it believes it has taken the lead. The companion’s dated bets are wired into this booklet’s chips section.

The 2015 parallel

Anyone who lived through machine learning’s enterprise adoption between 2014 and 2018 has seen this shape before. Statisticians and data scientists arrived from math and statistics backgrounds, fluent in modeling but uneven in software engineering. They built models in notebooks; the models worked in the notebook; the models did not ship. The gap was real and load-bearing rather than a fashionable complaint. The eventual resolution was a decade of work on MLOps, feature stores, model registries, and cross-functional teams pairing data scientists with software engineers and platform people. The capability was always there. The path from capability to production took the better part of ten years to build out.

The LLM gap has the same shape and is hitting enterprises hard right now. Capability has run ahead of the operational maturity to deploy it. Pilots multiply; production deployments lag. The MIT NANDA “95% zero ROI” figure has measurement issues, but even with conservative reframings the underlying message is correct: most enterprises haven’t finished the deployment side. The 61% of AI subscriptions Accenture identified as “underutilized due to poor integration” is the same story stated more carefully. The same talent and tooling gap. Same response: build the bridge.

What’s different this time

The 2015 parallel is the right scaffold but it isn’t a copy-paste. Two things are materially different, and they should reshape how teams budget the bridge.

First, far less data-pipeline work. The 2015 ML era spent enormous effort on data engineering: ETL, feature pipelines, training-serving skew, feature stores. LLMs invert most of that. They generate outputs from unstructured inputs rather than processing structured data; the data layer is retrieval and context assembly, not feature transformation; high-volume ETL is largely not the bottleneck. Teams that assume their LLM deployment needs an MLOps-shaped data team will mis-budget. The work is real but it sits elsewhere.

Second, far more testing and validation work. This is the part most enterprises systematically underestimate. An LLM can confabulate plausibly, an agent can take actions, output reaches end-users directly, and the damage potential of an undetected failure is qualitatively higher than “our regression test set drifted.” The work that was once 10–15% of MLOps spend (evaluation, monitoring, output review) becomes first-class infrastructure. Eval pipelines, red-teaming, calibration of human review, behavior change-management when a model upgrades: this is the deployment work itself. Teams that staff the bridge with the MLOps shape will discover the bridge is built wrong.

The benchmark-to-deployment gap, quantified

The same pattern shows up in every domain that has been measured carefully. GPT-4 achieved 92% diagnostic accuracy in controlled medical studies, but a meta-analysis across 83 studies found only 52.1% overall AI diagnostic accuracy in real-world settings, nearly a 40-point gap. (The two numbers come from different studies, models, and task designs, so don’t treat the subtraction as a measurement; the consistent shape, lab score far above field score, is the finding.) On the SWE-Lancer benchmark of real freelance coding tasks, even top models succeed only 26.2% of the time despite near-perfect HumanEval scores. On RE-Bench long-horizon tasks, AI systems score 4× higher than humans at 2 hours but humans outperform AI 2:1 at 32 hours. Deloitte’s 2026 enterprise survey found 20% of enterprises reporting AI-driven revenue, with two-thirds still stuck in pilot. Every one of these numbers describes a deployment problem, and none of them a capability problem.

Regulation as a secondary force raising the floor

Regulation plays a supporting role in this current rather than the headline, but it is a real second-order force, and one piece of news clarifies how to weight it. On May 7, 2026, EU negotiators reached provisional political agreement on the Digital Omnibus on AI: Annex III high-risk obligations are postponed from August 2, 2026 to December 2, 2027 (a 16-month deferral), and Annex I product-regulated high-risk obligations are deferred from August 2, 2027 to August 2, 2028. Watermarking and AI-content transparency shift by only three months, to December 2, 2026.

The delay should not reduce urgency for buyers. General-purpose AI model obligations under Articles 50–55 are unchanged and continue on the original schedule. Article 5 prohibitions are already in force. The Article 4 AI literacy obligation is already binding. Standards and guidance will still publish close to the new deadlines. The Code of Practice on synthetic content is expected to finalise in May or June 2026. What the omnibus moved was the most expensive, most operationally heavy obligations, precisely the ones tied to deployment of high-risk systems. The deployment gap is the headline; regulation is the floor underneath it, which the omnibus moved but did not remove.

Copyright runs in parallel. The Bartz v. Anthropic case produced a $1.5 billion class-wide settlement. The New York Times v. OpenAI multi-district litigation is still grinding through discovery: the expected spring 2026 summary judgment did not materialize; instead, a January 2026 ruling forced OpenAI to hand over a 20-million-conversation sample of ChatGPT logs, which will shape the fair-use fight. There are 56+ ongoing copyright lawsuits against AI companies. Every settlement raises the floor of data-provenance and due-diligence work required to deploy an LLM in production. Treat this as friction on the deployment side, not as a separate force.

What teams that bridge the gap actually look like

The 2015 resolution was cross-functional teams: data scientist plus software engineer plus platform engineer. The 2026 resolution looks similar in shape but reweighted: evaluation engineers, red-team specialists, and workflow designers become first-class roles. Less feature-store work; more behavioral testing. Less ETL; more output review. Less concept drift; more model upgrades that change personality. The teams that ship LLM features into production at scale in 2026 are the ones that have already staffed this shape. Most haven’t.

Two units

The conversation about AI capability is changing because the people who buy capability are not benchmark researchers. Employers do not care whether a model added two points on MMLU. They care about two numbers. First: how many hours an agent can work on something autonomously before a human needs to step in. Second: what an hour of that autonomous work costs, in API tokens or compute, compared to the loaded hourly rate of the person it would otherwise be done by. Those two numbers, multiplied, are the displacement math. The first is improving observably on a roughly four-month doubling cadence. The second is collapsing on the curve Section 2 covers. The product of the two is what will decide which work moves and which doesn’t.

Where the autonomy is today

The first unit is no longer hypothetical. The observable artifacts as of mid-2026: Claude Code routinely runs multi-hour autonomous coding sessions; Anthropic’s Computer Use lets an agent drive a desktop directly; Cursor (with Composer 2.5) and Windsurf (bundling Devin Cloud) sell agentic coding by the hour, not by the demo. Google demonstrated Antigravity 2.0 by having 93 sub-agents generate 2.6 billion tokens to build the core framework of an operating system in roughly 12 hours: theatre, yes, but the artifact existed. Gemini Spark, announced at I/O on May 19, is a personal agent that runs cloud-side 24/7 even when the user’s device is off. Anthropic’s “Dreaming” feature gives models memory consolidation across long-running work. None of these tools clear the 8-hour undisturbed-work threshold reliably yet, but several can sustain hours of useful autonomous work in narrow domains. That was not true a year ago.

The cost comparison

The displacement framing only works if the second unit lines up with the first. Take a representative knowledge-worker task that takes a senior practitioner about eight hours: a mid-complexity coding ticket with tests, or a structured analysis with document synthesis. Today’s math, using May 2026 prices and observed Claude Code token telemetry:

• Claude Opus 4.7 at $5 input / $25 output per million tokens. A moderately loaded autonomous coding agent burns on the order of 1M tokens per hour (typically ~700K input, ~300K output, with most input cached). At cached-input pricing, that lands at roughly $8 per agent-hour. An 8-hour autonomous run lands near $65–$80. A heavy multi-tool autonomous workload pushing 3–5M tokens per hour pushes per-hour cost into the $25–$45 range.
• Claude Sonnet 4.6 at $3 input / $15 output per million tokens. Same workload at moderate load: roughly $5 per agent-hour (the ~300K output tokens alone cost $4.50); an 8-hour run lands near $35–$45. Heavy load reaches $15–$25 per hour.
• Claude Fable 5 (July 2026 update) at $10 input / $50 output per million tokens, API-only post-recall. Same moderate load: roughly $16 per agent-hour (the ~300K output tokens alone cost $15); an 8-hour run lands near $130–$160. The premium is priced against measured capability; see “What the Fable 5 pricing does to the math” below.
• Senior developer comparator: in the US, loaded hourly cost typically lands at $130–$245 per hour (base $110–$175 plus 20–40% loaded for benefits, taxes, overhead). In Western Europe the equivalent runs €100–€150 fully loaded; in CEE roughly half that. Per 8-hour day, that’s $1,000–$2,000 US / €800–€1,200 Western EU.

The raw ratio, before any overhead, is striking: an autonomous Opus 4.7 hour costs roughly 15–30× less than the senior US developer hour it would replace; an autonomous Sonnet 4.6 hour costs roughly 30–50× less; even a Fable 5 hour, the most expensive agent-hour on the market, comes in roughly 8–15× below the human comparator. Most observers stop here, get excited, and reach the wrong conclusion. The honest number adds review and retry overhead: every autonomous hour today realistically needs roughly 20–40 minutes of human supervision, evaluation runs, and retry cycles before the work ships. That overhead compresses the effective gap into something more like 3–10× cheaper for the right workflow, and to break-even or worse for workflows where the model still gets stuck.

Section 2 explains why the underlying token gap closes faster than people expect: frontier-equivalent inference costs fell ~99% in two years and the April 2026 open-weight wave (DeepSeek V4-Pro at $0.28/$2.48 per MTok) is another factor of 10 below paid frontier. The crossover for any specific workflow depends mostly on two things: how much human supervision overhead it still needs, and what the loaded-cost comparator actually is in your geography. Pick one workflow your team actually does. Estimate both numbers. Track the ratio quarterly. That ratio, more than any benchmark, will tell you when displacement becomes economic.

The METR data point

One supporting data point is worth keeping in view. METR (Model Evaluation & Threat Research) publishes a measured benchmark called Time Horizon: the duration of human work a model can complete with 50% reliability. Their 2019–2025 dataset showed the frontier doubling roughly every 7 months. Their TH 1.1 update (January 2026) expanded the task suite by a third and doubled the number of 8-hour-plus tasks; measured from 2024 onward, the doubling time comes out near 3 months (89 days). The anchors move fast: Claude 3.7 Sonnet (early 2025) sat around a 50-minute horizon, while METR’s February–March 2026 pilot with the frontier labs reported the strongest agents near or beyond the reliable measurement range of the benchmark itself. If the ~3–4-month doubling holds, frontier models cross the 8-hour threshold by late 2026 or early 2027; the ruler may run out before the calendar does. If the older 7-month cadence reasserts, that slips toward 2028. Either way, the trajectory is the input to the displacement curve, not the headline number.

The Remote Labor Index: displacement gets a leaderboard

As of July 2026, this current has the instrument it was missing. The Remote Labor Index (Center for AI Safety and Scale AI) is built from real commissioned freelance projects: more than 6,000 hours of professional work worth over $140,000, across eight domains from CAD and architecture to data analysis and animation. Its metric is exactly this chapter’s question in benchmark form: the automation rate, the share of projects where the AI agent’s deliverable is judged as good as the paid professional’s. Unlike SWE-bench or MMLU, the units are dollars and deliverables, not points.

The curve so far: 2.5% at launch in late 2025; 4.17% by the spring (Opus 4.6 running in an agentic scaffold); then on July 1, 2026, CAIS published the first Mythos-class result: Fable 5 at 16.1% (measured on 218 of 240 projects before the access restrictions landed), roughly double Opus 4.8’s 8.3%, with GPT-5.5 at 6.3%. The frontier more than quadrupled in under eight months. Two readings of that number are simultaneously true, and both matter. Read pessimistically: 84% of real, paid remote-work projects still can’t be automated end-to-end, and even headline deliverables often wouldn’t ship as finished client work; that is the lab-to-production gap of Current 5, measured in invoices. Read as a trajectory: this is the METR doubling cadence showing up in a dollar-denominated instrument, and it converts the “hours” unit of this chapter into something an employer can put in a spreadsheet.

What the Fable 5 pricing does to the math

The June frontier release also moved the second unit, upward, which is why this section’s numbers need re-basing. Fable 5 launched at $10 input / $50 output per million tokens, double Opus 4.7 on both meters, and its post-recall availability is API-only with a strong moderation layer. The same moderate autonomous load (~1M tokens per hour) lands at roughly $16 per agent-hour; the ~300K output tokens alone cost $15. But price both sides of the trade: at roughly 2× the price it posts roughly 2× the RLI automation rate of Opus 4.8, so capability per dollar at the very top held flat-to-better even as the absolute price rose. This is Current 2’s two-lane market seen from the buyer’s chair: the frontier lane charges a premium that, for now, buys measurably more finished work per dollar on the hardest tasks, while one tier down, the commodity lane keeps collapsing toward hardware cost.

What this implies for capex

If “hours of autonomous work” is the right capability axis, the capex picture from Section 1 changes shape. The training portion of hyperscaler spend (the largest pre-training runs) becomes harder to justify on its own; smaller models with strong post-training, RL, and tool use can match the capabilities of larger ones for many workflows. The inference portion, projected at roughly 75% of AI compute demand by 2030 reported, becomes more justified, because every long-running agent is a multi-token, multi-call, often multi-hour inference workload. Reasoning models with extended thinking can use 10–100× more tokens per task than chat-style models. The capex stays justified; what shifts is its allocation between training and serving.

The bit-atom mismatch

Every timescale in the capex decoder is really a physical timescale wearing a financial costume. Building a leading-edge fab takes three to five years. A gigawatt-scale grid interconnection can take most of a decade in congested regions. A data center takes 12–24 months if the power is already there. A frontier training run takes months. The decoder’s clean arithmetic (money in year N, capability in year N+2 to N+4) holds only as long as the slowest of those clocks doesn’t slip. This current watches the clocks.

Power: the queue is the moat

The Lawrence Berkeley National Laboratory’s queue tracker counts roughly 2,300 gigawatts of generation and storage waiting for US grid interconnection, nearly twice the country’s entire installed capacity, with typical waits measured in years. Meanwhile the demand side compounds: the LBNL/DOE energy report has data centers at 4.4% of US electricity in 2023, projected to 6.7–12% by 2028, and 451 Research has data-center grid demand more than doubling from 61.8 GW in 2025 to 134 GW by 2030. The market has already priced the collision. PJM, the grid operator for the world’s densest data-center corridor, has now cleared three consecutive capacity auctions at the regulatory price cap; the December 2025 auction settled at $333.44/MW-day and would have cleared near $530 without the cap, with the market monitor attributing data centers as the primary driver of roughly 40% of that auction’s $16.4 billion cost. This is why hyperscalers are buying power plants rather than power: the Three Mile Island Unit 1 restart (now the Crane Clean Energy Center) exists because Microsoft signed a 20-year deal for all 835 MW of it, and it is tracking toward a second-half-2027 restart, a year ahead of the original plan. The first US nuclear reactor brought back from retirement specifically to feed AI is a fact the 2015 version of this industry would have found unbelievable.

Chips: one island, one packager, three memory vendors

The silicon side concentrates even harder. TSMC fabricates roughly 90% of the world’s most advanced logic chips, and its advanced nodes (the ones AI accelerators use) now account for about 74% of its wafer revenue, with price increases landing across all of them. The bottleneck within the bottleneck is advanced packaging: CoWoS capacity is fully booked with lead times of 52–78 weeks, demand approaching a million wafers in 2026, and Nvidia alone holding an estimated 60% of it. High-bandwidth memory is a three-vendor oligopoly in which SK Hynix holds roughly 62%, and HBM3E is effectively sold out for all of 2026. Read those three sentences against the decoder: when the inputs to capability are pre-sold 12–18 months out, capex stops being a dial you can turn and becomes a queue position you defend.

The politics of power arrived

Through 2025 this current was an engineering story. In 2026 it grew a political layer, fast. More than 300 data-center bills were filed in over 30 states in the first six weeks of the year (up from about 200 bills in all of 2025), with at least a dozen states proposing outright moratoria. Maine’s legislature actually passed the first statewide ban (LD 307, blocking data centers over 20 MW); Governor Mills vetoed it in April and the override failed by seven votes. New York’s legislature passed a one-year moratorium of its own, awaiting signature as this edition went to press. The pattern to watch is not any single bill but the direction: retail electricity bills are becoming the political transmission mechanism, since IEEFA attributes 63% of the price increase in PJM’s 2025/26 auction, some $9.3 billion recoverable from ratepayers, to data centers, and voters pay retail bills. In parallel, the federal layer began treating compute itself as a controlled substance: rules effective January 15, 2026 codified case-by-case export review for advanced AI chips and imposed a 25% tariff on advanced AI chips not destined for the American supply chain. The companion booklet’s “bullion is compute” chapter argues this is exactly where mercantilist control naturally lands, because concentrated compute is the one input that cannot leak.

Why this is a current, not a footnote to Current 1

Current 1 asks whether the money keeps being right. This current asks whether the money can be spent. Those are different failure modes with different signatures: Current 1 fails through disappointing models, Current 3 through disappearing funding, and Current 7 through slipping dates, regional scarcity, and price premiums that have nothing to do with capability. It was promoted from background assumption to full current in this edition because 2026 is the year its trigger surface became observable: capacity auctions at caps, moratorium bills in a dozen statehouses, packaging lead times crossing a year and a half. The atoms got political.

Strip the statistics away and this current asks a plain question: when AI starts doing real work, what do the people whose work it was, and the politicians who represent them, do about it? Current 6 watches whether the automation math works. This current watches the human response to it: workers organizing, companies hiding the ball, vendors quietly pre-paying their political bills, legislators hunting for something to tax. Four scenes carry the story.

The canaries stopped singing

The cleanest evidence that displacement is real comes from the Stanford Digital Economy Lab’s “Canaries in the Coal Mine” study, run on ADP payroll data covering millions of workers. Its core finding is simple: AI absorbs tasks before it absorbs jobs, and entry-level tasks first. Since generative AI arrived, employment of the youngest workers (22–25) in the most AI-exposed occupations has fallen roughly 16%, while older workers in the same occupations at the same firms grew. Read that again: same firm, same job title, opposite fates, depending on whether your daily work was the entry-level slice. The follow-up dashboard shows the decline still accelerating through spring 2026, and the Stanford AI Index adds the sharpest sectoral cut: employment of young software developers down nearly 20% since 2024. This is Current 6’s economics arriving in payroll data, youngest cohort first, exactly where the hours-and-dollars math said it would bite.

Attribution laundering

Now put that evidence next to the official record, because the two tell opposite stories. New York became the first state to require employers to say whether AI caused a mass layoff. Eleven months in, zero of 162 filings have said yes. Over the same stretch, the running press tally of publicly AI-attributed job cuts (Challenger, Gray & Christmas) reached 87,714 in the first five months of 2026, more than in all of 2025. Both numbers can be true at once because attribution is a choice: a layoff filed as “restructuring” is legally cleaner and reputationally cheaper than one filed as “automation.” Amazon cutting roughly 30,000 corporate roles while spending $125 billion on AI capex, with reported internal plans to avoid some 600,000 future hires through robotics and AI, is the same move at its largest scale. The planning consequence is unusual: official statistics will systematically understate this current, so its triggers have to watch other venues (filing gaps, union demands, private funds) rather than the headline numbers.

The tax question arrives before the tax

The most telling 2026 development is who moved first. On April 6, OpenAI published economic-policy proposals that include a potential robot tax, a “Public Wealth Fund” giving Americans an automatic stake in AI companies and infrastructure, and a subsidized four-day workweek. Sit with that for a second: the company building the technology is designing its own redistribution mechanism, unprompted. In June, the RAISE US workforce fund launched with more than $500 million raised toward a $1 billion goal, anchored by Amazon, Anthropic, Microsoft, and the OpenAI Foundation. Meanwhile the definitional problem keeps actual legislation stuck: as Bloomberg Law’s survey of tax authorities put it, governments want to tax AI but cannot yet define the unit. Tokens? Agent-hours? Compute? Displaced heads? When an industry starts pre-paying a political bill nobody has formally presented, it is telling you what it expects this current to deliver.

Labor’s playbook already exists

The response side is not starting from zero. SAG-AFTRA’s 2023 AI provisions, the Las Vegas Culinary Union’s technology-severance terms ($2,000 per year worked for tech-displaced workers), and the longshoremen’s January 2025 contract restricting port automation all predate the agent era. In 2026 the playbook entered knowledge work: a one-day strike at ProPublica in April with AI use as a central issue, a New York Times union letter calling the paper’s AI policies inadequate the day before, an AP union unfair-labor-practice complaint over AI the day before that, and a December 2025 arbitration win at Politico for launching AI products without consulting the union. AI clauses are becoming a normal bargaining demand. None of this is yet a 10,000-worker action, but the templates are written, tested, and circulating.

Why this is a current, not a chapter of Current 6

Current 6 measures the economics; this current tracks the response function, and response functions are discontinuous where economics are smooth. The RLI is the hinge between them: the same instrument that tells employers the math is starting to work tells legislators it is working. Note the bookkeeping implication, because it matters for the synthesis: when one RLI publication lands in both chapters’ trigger logs, that is one event wearing two jackets. It counts once under the one-event rule, not as a fired pair.